In October 2024 OpenAI published Swarm, an open-source framework on GitHub for orchestrating multiple agents. OpenAI was deliberate that Swarm was educational rather than production software - a way to explore ergonomic, lightweight multi-agent coordination and make it controllable and testable, not a supported product.
Swarm reduced multi-agent orchestration to two primitives. An agent bundles a set of instructions with a set of functions it can call. A handoff is the act of one agent transferring the ongoing conversation to another, expressed simply by having a function return the next agent to run. With these two ideas a developer can route a conversation - say from a triage agent to a billing agent to a refunds agent - without a heavy framework. Swarm runs almost entirely on the client and is stateless between calls, and OpenAI noted its agents are unrelated to the Assistants in the Assistants API.
Swarm mattered because it signaled OpenAI’s own framing of how to build agent systems - small, composable agents that pass control among themselves - at a moment when LangGraph, CrewAI, and AutoGen were competing to define the pattern. OpenAI later replaced Swarm with the production-oriented Agents SDK, describing it as the evolution of the ideas Swarm prototyped.